Abstract: Scene classification is significant for scene imagery semantic interpretation, which is one of the main research topics of remote sensing filed. The traditional image classification methods usually extract the insufficient features with much time in training its classifier. In addition, it relies on large number of manually labeled data. Aiming at solving these problems, this paper proposes a scene image classification model based on multi-feature and constrained extreme learning machine（CELM）. There are two main processes of the proposed method. Firstly, three pre-trained convolutional neural networks with different architectures are fine-tuned using the experimental dataset. Secondly, the features extracted are fused and feed to CELM classifier, obtaining the category label. SIRI-WHU, WHU-RS19 and UC-Merced datasets are taken as the experimental datasets, the experiments on pre-trained CNNs, single features and traditional classifiers show that the proposed method produces the comparable classification performance, with the accuracy of 99.25%, 98.26% and 97.70%.

Key words: scene classification, multi-feature fusion, constrained extreme learning machine

摘要： 场景分类对于场景图像的语义解译至关重要，是遥感领域近期的主要研究课题之一。针对大部分图像分类方法中提取的特征结构单一，依赖于大量人工标记的数据以及分类器的训练过程缓慢等问题，提出了一种基于多特征融合与约束极限学习机（constrained extreme learning machines，CELM）的场景图像分类方法。该方法采用三种不同结构的预训练卷积神经网络，利用特定数据集对其进行微调，将微调后网络提取到的三种特征进行融合并送入CELM分类器进行分类，最终得到图像的类别标签。以SIRI-WHU、WHU-RS19与UC-Merced数据集作为实验数据集，在预训练卷积神经网络、单一特征和传统分类器上进行的对比实验表明，基于多特征融合与CELM相结合的方法产生了较好的分类效果，三种数据集上的总分类精度分别高达99.25%、98.26%与97.70%。

关键词: 场景分类, 多特征融合, 约束极限学习机